10k

Vision Transformers (ViTs)

Let's journey through the fascinating evolution from text-based Large Language Models (LLMs) like BERT to the visually astute world of Vision Transformers (ViTs). It's like going from reading a book to understanding a picture book with the same depth and nuance.

How LLMs and Vision Transformers (ViTs) Connect

Imagine LLMs as being super smart at "reading between the lines" of text. They can understand context, emotion, and even the unspoken bits of language. Now, what if we could teach a computer to "see" images with the same depth? Enter Vision Transformers! ViTs chop up images into smaller pieces, almost like how we look at words in a sentence, and use the magic of attention mechanisms (yep, the same ones from LLMs) to see the bigger picture.

Why is this so powerful

When we blend the intelligence of LLMs with the visual savvy of ViTs, we get something extraordinary: multimodal LLMs (MLLMs) – something we covered in our early bonus modules. An example of such an application is ChatGPT Plus which can generate and comprehend both images as well as texts.

Lecture on Vision Transformers

To explore this exciting intersection further, here is a recorded lecture by Dr. Vijay Srinivas Agneeswaran (Sr. Director and ML Research Leader at Microsoft), who will provide an in-depth look at the latest advancements in computer vision, particularly Vision Transformers. This session is a unique opportunity to gain insights into the transition from traditional convolutional neural networks to transformers in the visual domain, and how these technologies are contributing to the evolution of LLMs.

Moderated by Mudit Srivastava from Pathway, this session starts with a brief introduction of Pathway by Claire Nouet (COO, Pathway), followed by an introduction and a deep dive with Vijay.

Key Takeaways

  • An understanding of the transition from traditional computer vision techniques to the use of pre-trained transformers.
  • Insights into how Vision Transformers and LLMs can work together to enhance AI's capabilities in tasks involving both text and images.
  • Detailed knowledge of Scattering Vision Transformers (SVT), their development, and their application in real-world tasks.

More about the educator, Vijay S Agneeswaran

Dr. Vijay Srinivas Agneeswaran, Sr. Director and ML Research Leader at Microsoft, brings over two decades of expertise in AI, machine learning, and data science. Holding a Ph.D. from IIT Madras and a postdoctoral from EPFL, his specializations include computer vision, efficient transformers, and large language models. At Microsoft, he has led pioneering research in AI for C+AI data and developed spectral transformers for computer vision, showcased at NeurIPS 2023. He is a champion of responsible AI, ensuring compliance for nearly 50 AI models, and has led teams in organizations like Walmart Global Tech, Oracle, and Cognizant, attesting to his significant industry impact. Dr. Agneeswaran also holds five US patents and is a prolific contributor to tech conferences and publications. You can find him on Twitter here.